Upload folder using huggingface_hub
Browse files
main/regional_prompting_stable_diffusion.py
CHANGED
@@ -1,14 +1,43 @@
|
|
|
|
1 |
import math
|
2 |
-
from typing import Dict, Optional
|
3 |
|
4 |
import torch
|
5 |
import torchvision.transforms.functional as FF
|
6 |
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
|
7 |
|
8 |
-
from diffusers import
|
|
|
|
|
|
|
|
|
|
|
|
|
9 |
from diffusers.models import AutoencoderKL, UNet2DConditionModel
|
|
|
|
|
|
|
10 |
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
|
11 |
from diffusers.schedulers import KarrasDiffusionSchedulers
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
|
13 |
|
14 |
try:
|
@@ -21,7 +50,14 @@ KCOMM = "ADDCOMM"
|
|
21 |
KBRK = "BREAK"
|
22 |
|
23 |
|
24 |
-
class RegionalPromptingStableDiffusionPipeline(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
r"""
|
26 |
Args for Regional Prompting Pipeline:
|
27 |
rp_args:dict
|
@@ -78,17 +114,7 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
|
|
78 |
image_encoder: CLIPVisionModelWithProjection = None,
|
79 |
requires_safety_checker: bool = True,
|
80 |
):
|
81 |
-
super().__init__(
|
82 |
-
vae,
|
83 |
-
text_encoder,
|
84 |
-
tokenizer,
|
85 |
-
unet,
|
86 |
-
scheduler,
|
87 |
-
safety_checker,
|
88 |
-
feature_extractor,
|
89 |
-
image_encoder,
|
90 |
-
requires_safety_checker,
|
91 |
-
)
|
92 |
self.register_modules(
|
93 |
vae=vae,
|
94 |
text_encoder=text_encoder,
|
@@ -99,6 +125,17 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
|
|
99 |
feature_extractor=feature_extractor,
|
100 |
image_encoder=image_encoder,
|
101 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
102 |
|
103 |
@torch.no_grad()
|
104 |
def __call__(
|
@@ -413,7 +450,7 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
|
|
413 |
|
414 |
hook_forwards(self.unet)
|
415 |
|
416 |
-
output =
|
417 |
prompt=prompt,
|
418 |
prompt_embeds=embs,
|
419 |
negative_prompt=negative_prompt,
|
@@ -449,6 +486,909 @@ class RegionalPromptingStableDiffusionPipeline(StableDiffusionPipeline):
|
|
449 |
|
450 |
return output
|
451 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
452 |
|
453 |
### Make prompt list for each regions
|
454 |
def promptsmaker(prompts, batch):
|
@@ -663,3 +1603,88 @@ def scaled_dot_product_attention(
|
|
663 |
get_attn_maps(self, attn_weight)
|
664 |
attn_weight = torch.dropout(attn_weight, dropout_p, train=True)
|
665 |
return attn_weight @ value
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import inspect
|
2 |
import math
|
3 |
+
from typing import Any, Callable, Dict, List, Optional, Union
|
4 |
|
5 |
import torch
|
6 |
import torchvision.transforms.functional as FF
|
7 |
from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
|
8 |
|
9 |
+
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
|
10 |
+
from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
|
11 |
+
from diffusers.loaders import StableDiffusionLoraLoaderMixin
|
12 |
+
from diffusers.loaders.ip_adapter import IPAdapterMixin
|
13 |
+
from diffusers.loaders.lora_pipeline import LoraLoaderMixin
|
14 |
+
from diffusers.loaders.single_file import FromSingleFileMixin
|
15 |
+
from diffusers.loaders.textual_inversion import TextualInversionLoaderMixin
|
16 |
from diffusers.models import AutoencoderKL, UNet2DConditionModel
|
17 |
+
from diffusers.models.lora import adjust_lora_scale_text_encoder
|
18 |
+
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
19 |
+
from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
|
20 |
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
|
21 |
from diffusers.schedulers import KarrasDiffusionSchedulers
|
22 |
+
from diffusers.utils import (
|
23 |
+
USE_PEFT_BACKEND,
|
24 |
+
deprecate,
|
25 |
+
is_torch_xla_available,
|
26 |
+
logging,
|
27 |
+
scale_lora_layers,
|
28 |
+
unscale_lora_layers,
|
29 |
+
)
|
30 |
+
from diffusers.utils.torch_utils import randn_tensor
|
31 |
+
|
32 |
+
|
33 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
34 |
+
|
35 |
+
if is_torch_xla_available():
|
36 |
+
import torch_xla.core.xla_model as xm
|
37 |
+
|
38 |
+
XLA_AVAILABLE = True
|
39 |
+
else:
|
40 |
+
XLA_AVAILABLE = False
|
41 |
|
42 |
|
43 |
try:
|
|
|
50 |
KBRK = "BREAK"
|
51 |
|
52 |
|
53 |
+
class RegionalPromptingStableDiffusionPipeline(
|
54 |
+
DiffusionPipeline,
|
55 |
+
TextualInversionLoaderMixin,
|
56 |
+
LoraLoaderMixin,
|
57 |
+
IPAdapterMixin,
|
58 |
+
FromSingleFileMixin,
|
59 |
+
StableDiffusionLoraLoaderMixin,
|
60 |
+
):
|
61 |
r"""
|
62 |
Args for Regional Prompting Pipeline:
|
63 |
rp_args:dict
|
|
|
114 |
image_encoder: CLIPVisionModelWithProjection = None,
|
115 |
requires_safety_checker: bool = True,
|
116 |
):
|
117 |
+
super().__init__()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
118 |
self.register_modules(
|
119 |
vae=vae,
|
120 |
text_encoder=text_encoder,
|
|
|
125 |
feature_extractor=feature_extractor,
|
126 |
image_encoder=image_encoder,
|
127 |
)
|
128 |
+
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
|
129 |
+
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
|
130 |
+
self.register_to_config(requires_safety_checker=requires_safety_checker)
|
131 |
+
|
132 |
+
# Initialize additional properties needed for DiffusionPipeline
|
133 |
+
self._num_timesteps = None
|
134 |
+
self._interrupt = False
|
135 |
+
self._guidance_scale = 7.5
|
136 |
+
self._guidance_rescale = 0.0
|
137 |
+
self._clip_skip = None
|
138 |
+
self._cross_attention_kwargs = None
|
139 |
|
140 |
@torch.no_grad()
|
141 |
def __call__(
|
|
|
450 |
|
451 |
hook_forwards(self.unet)
|
452 |
|
453 |
+
output = self.stable_diffusion_call(
|
454 |
prompt=prompt,
|
455 |
prompt_embeds=embs,
|
456 |
negative_prompt=negative_prompt,
|
|
|
486 |
|
487 |
return output
|
488 |
|
489 |
+
# copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion
|
490 |
+
def prepare_extra_step_kwargs(self, generator, eta):
|
491 |
+
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
|
492 |
+
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
|
493 |
+
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
|
494 |
+
# and should be between [0, 1]
|
495 |
+
|
496 |
+
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
497 |
+
extra_step_kwargs = {}
|
498 |
+
if accepts_eta:
|
499 |
+
extra_step_kwargs["eta"] = eta
|
500 |
+
|
501 |
+
# check if the scheduler accepts generator
|
502 |
+
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
|
503 |
+
if accepts_generator:
|
504 |
+
extra_step_kwargs["generator"] = generator
|
505 |
+
return extra_step_kwargs
|
506 |
+
|
507 |
+
# copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion
|
508 |
+
def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
|
509 |
+
shape = (
|
510 |
+
batch_size,
|
511 |
+
num_channels_latents,
|
512 |
+
int(height) // self.vae_scale_factor,
|
513 |
+
int(width) // self.vae_scale_factor,
|
514 |
+
)
|
515 |
+
if isinstance(generator, list) and len(generator) != batch_size:
|
516 |
+
raise ValueError(
|
517 |
+
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
|
518 |
+
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
|
519 |
+
)
|
520 |
+
|
521 |
+
if latents is None:
|
522 |
+
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
523 |
+
else:
|
524 |
+
latents = latents.to(device)
|
525 |
+
|
526 |
+
# scale the initial noise by the standard deviation required by the scheduler
|
527 |
+
latents = latents * self.scheduler.init_noise_sigma
|
528 |
+
return latents
|
529 |
+
|
530 |
+
# copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion
|
531 |
+
def encode_prompt(
|
532 |
+
self,
|
533 |
+
prompt,
|
534 |
+
device,
|
535 |
+
num_images_per_prompt,
|
536 |
+
do_classifier_free_guidance,
|
537 |
+
negative_prompt=None,
|
538 |
+
prompt_embeds: Optional[torch.Tensor] = None,
|
539 |
+
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
540 |
+
lora_scale: Optional[float] = None,
|
541 |
+
clip_skip: Optional[int] = None,
|
542 |
+
):
|
543 |
+
r"""
|
544 |
+
Encodes the prompt into text encoder hidden states.
|
545 |
+
|
546 |
+
Args:
|
547 |
+
prompt (`str` or `List[str]`, *optional*):
|
548 |
+
prompt to be encoded
|
549 |
+
device: (`torch.device`):
|
550 |
+
torch device
|
551 |
+
num_images_per_prompt (`int`):
|
552 |
+
number of images that should be generated per prompt
|
553 |
+
do_classifier_free_guidance (`bool`):
|
554 |
+
whether to use classifier free guidance or not
|
555 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
556 |
+
The prompt or prompts not to guide the image generation. If not defined, one has to pass
|
557 |
+
`negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
|
558 |
+
less than `1`).
|
559 |
+
prompt_embeds (`torch.Tensor`, *optional*):
|
560 |
+
Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
|
561 |
+
provided, text embeddings will be generated from `prompt` input argument.
|
562 |
+
negative_prompt_embeds (`torch.Tensor`, *optional*):
|
563 |
+
Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
|
564 |
+
weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
|
565 |
+
argument.
|
566 |
+
lora_scale (`float`, *optional*):
|
567 |
+
A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
|
568 |
+
clip_skip (`int`, *optional*):
|
569 |
+
Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
|
570 |
+
the output of the pre-final layer will be used for computing the prompt embeddings.
|
571 |
+
"""
|
572 |
+
# set lora scale so that monkey patched LoRA
|
573 |
+
# function of text encoder can correctly access it
|
574 |
+
if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin):
|
575 |
+
self._lora_scale = lora_scale
|
576 |
+
|
577 |
+
# dynamically adjust the LoRA scale
|
578 |
+
if not USE_PEFT_BACKEND:
|
579 |
+
adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
|
580 |
+
else:
|
581 |
+
scale_lora_layers(self.text_encoder, lora_scale)
|
582 |
+
|
583 |
+
if prompt is not None and isinstance(prompt, str):
|
584 |
+
batch_size = 1
|
585 |
+
elif prompt is not None and isinstance(prompt, list):
|
586 |
+
batch_size = len(prompt)
|
587 |
+
else:
|
588 |
+
batch_size = prompt_embeds.shape[0]
|
589 |
+
|
590 |
+
if prompt_embeds is None:
|
591 |
+
# textual inversion: process multi-vector tokens if necessary
|
592 |
+
if isinstance(self, TextualInversionLoaderMixin):
|
593 |
+
prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
|
594 |
+
|
595 |
+
text_inputs = self.tokenizer(
|
596 |
+
prompt,
|
597 |
+
padding="max_length",
|
598 |
+
max_length=self.tokenizer.model_max_length,
|
599 |
+
truncation=True,
|
600 |
+
return_tensors="pt",
|
601 |
+
)
|
602 |
+
text_input_ids = text_inputs.input_ids
|
603 |
+
untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
|
604 |
+
|
605 |
+
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
|
606 |
+
text_input_ids, untruncated_ids
|
607 |
+
):
|
608 |
+
removed_text = self.tokenizer.batch_decode(
|
609 |
+
untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
|
610 |
+
)
|
611 |
+
logger.warning(
|
612 |
+
"The following part of your input was truncated because CLIP can only handle sequences up to"
|
613 |
+
f" {self.tokenizer.model_max_length} tokens: {removed_text}"
|
614 |
+
)
|
615 |
+
|
616 |
+
if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
|
617 |
+
attention_mask = text_inputs.attention_mask.to(device)
|
618 |
+
else:
|
619 |
+
attention_mask = None
|
620 |
+
|
621 |
+
if clip_skip is None:
|
622 |
+
prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
|
623 |
+
prompt_embeds = prompt_embeds[0]
|
624 |
+
else:
|
625 |
+
prompt_embeds = self.text_encoder(
|
626 |
+
text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
|
627 |
+
)
|
628 |
+
# Access the `hidden_states` first, that contains a tuple of
|
629 |
+
# all the hidden states from the encoder layers. Then index into
|
630 |
+
# the tuple to access the hidden states from the desired layer.
|
631 |
+
prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
|
632 |
+
# We also need to apply the final LayerNorm here to not mess with the
|
633 |
+
# representations. The `last_hidden_states` that we typically use for
|
634 |
+
# obtaining the final prompt representations passes through the LayerNorm
|
635 |
+
# layer.
|
636 |
+
prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
|
637 |
+
|
638 |
+
if self.text_encoder is not None:
|
639 |
+
prompt_embeds_dtype = self.text_encoder.dtype
|
640 |
+
elif self.unet is not None:
|
641 |
+
prompt_embeds_dtype = self.unet.dtype
|
642 |
+
else:
|
643 |
+
prompt_embeds_dtype = prompt_embeds.dtype
|
644 |
+
|
645 |
+
prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
|
646 |
+
|
647 |
+
bs_embed, seq_len, _ = prompt_embeds.shape
|
648 |
+
# duplicate text embeddings for each generation per prompt, using mps friendly method
|
649 |
+
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
650 |
+
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
651 |
+
|
652 |
+
# get unconditional embeddings for classifier free guidance
|
653 |
+
if do_classifier_free_guidance and negative_prompt_embeds is None:
|
654 |
+
uncond_tokens: List[str]
|
655 |
+
if negative_prompt is None:
|
656 |
+
uncond_tokens = [""] * batch_size
|
657 |
+
elif prompt is not None and type(prompt) is not type(negative_prompt):
|
658 |
+
raise TypeError(
|
659 |
+
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
660 |
+
f" {type(prompt)}."
|
661 |
+
)
|
662 |
+
elif isinstance(negative_prompt, str):
|
663 |
+
uncond_tokens = [negative_prompt]
|
664 |
+
elif batch_size != len(negative_prompt):
|
665 |
+
raise ValueError(
|
666 |
+
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
667 |
+
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
668 |
+
" the batch size of `prompt`."
|
669 |
+
)
|
670 |
+
else:
|
671 |
+
uncond_tokens = negative_prompt
|
672 |
+
|
673 |
+
# textual inversion: process multi-vector tokens if necessary
|
674 |
+
if isinstance(self, TextualInversionLoaderMixin):
|
675 |
+
uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
|
676 |
+
|
677 |
+
max_length = prompt_embeds.shape[1]
|
678 |
+
uncond_input = self.tokenizer(
|
679 |
+
uncond_tokens,
|
680 |
+
padding="max_length",
|
681 |
+
max_length=max_length,
|
682 |
+
truncation=True,
|
683 |
+
return_tensors="pt",
|
684 |
+
)
|
685 |
+
|
686 |
+
if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
|
687 |
+
attention_mask = uncond_input.attention_mask.to(device)
|
688 |
+
else:
|
689 |
+
attention_mask = None
|
690 |
+
|
691 |
+
negative_prompt_embeds = self.text_encoder(
|
692 |
+
uncond_input.input_ids.to(device),
|
693 |
+
attention_mask=attention_mask,
|
694 |
+
)
|
695 |
+
negative_prompt_embeds = negative_prompt_embeds[0]
|
696 |
+
|
697 |
+
if do_classifier_free_guidance:
|
698 |
+
# duplicate unconditional embeddings for each generation per prompt, using mps friendly method
|
699 |
+
seq_len = negative_prompt_embeds.shape[1]
|
700 |
+
|
701 |
+
negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
|
702 |
+
|
703 |
+
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
704 |
+
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
|
705 |
+
|
706 |
+
if self.text_encoder is not None:
|
707 |
+
if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND:
|
708 |
+
# Retrieve the original scale by scaling back the LoRA layers
|
709 |
+
unscale_lora_layers(self.text_encoder, lora_scale)
|
710 |
+
|
711 |
+
return prompt_embeds, negative_prompt_embeds
|
712 |
+
|
713 |
+
# copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion
|
714 |
+
def check_inputs(
|
715 |
+
self,
|
716 |
+
prompt,
|
717 |
+
height,
|
718 |
+
width,
|
719 |
+
callback_steps,
|
720 |
+
negative_prompt=None,
|
721 |
+
prompt_embeds=None,
|
722 |
+
negative_prompt_embeds=None,
|
723 |
+
ip_adapter_image=None,
|
724 |
+
ip_adapter_image_embeds=None,
|
725 |
+
callback_on_step_end_tensor_inputs=None,
|
726 |
+
):
|
727 |
+
if height % 8 != 0 or width % 8 != 0:
|
728 |
+
raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
|
729 |
+
|
730 |
+
if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
|
731 |
+
raise ValueError(
|
732 |
+
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
|
733 |
+
f" {type(callback_steps)}."
|
734 |
+
)
|
735 |
+
if callback_on_step_end_tensor_inputs is not None and not all(
|
736 |
+
k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
|
737 |
+
):
|
738 |
+
raise ValueError(
|
739 |
+
f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
|
740 |
+
)
|
741 |
+
|
742 |
+
if prompt is not None and prompt_embeds is not None:
|
743 |
+
raise ValueError(
|
744 |
+
f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
|
745 |
+
" only forward one of the two."
|
746 |
+
)
|
747 |
+
elif prompt is None and prompt_embeds is None:
|
748 |
+
raise ValueError(
|
749 |
+
"Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
|
750 |
+
)
|
751 |
+
elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
|
752 |
+
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
|
753 |
+
|
754 |
+
if negative_prompt is not None and negative_prompt_embeds is not None:
|
755 |
+
raise ValueError(
|
756 |
+
f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
|
757 |
+
f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
|
758 |
+
)
|
759 |
+
|
760 |
+
if prompt_embeds is not None and negative_prompt_embeds is not None:
|
761 |
+
if prompt_embeds.shape != negative_prompt_embeds.shape:
|
762 |
+
raise ValueError(
|
763 |
+
"`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
|
764 |
+
f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
|
765 |
+
f" {negative_prompt_embeds.shape}."
|
766 |
+
)
|
767 |
+
|
768 |
+
if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
|
769 |
+
raise ValueError(
|
770 |
+
"Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
|
771 |
+
)
|
772 |
+
|
773 |
+
if ip_adapter_image_embeds is not None:
|
774 |
+
if not isinstance(ip_adapter_image_embeds, list):
|
775 |
+
raise ValueError(
|
776 |
+
f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
|
777 |
+
)
|
778 |
+
elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
|
779 |
+
raise ValueError(
|
780 |
+
f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
|
781 |
+
)
|
782 |
+
|
783 |
+
# copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion
|
784 |
+
@torch.no_grad()
|
785 |
+
def stable_diffusion_call(
|
786 |
+
self,
|
787 |
+
prompt: Union[str, List[str]] = None,
|
788 |
+
height: Optional[int] = None,
|
789 |
+
width: Optional[int] = None,
|
790 |
+
num_inference_steps: int = 50,
|
791 |
+
timesteps: List[int] = None,
|
792 |
+
sigmas: List[float] = None,
|
793 |
+
guidance_scale: float = 7.5,
|
794 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
795 |
+
num_images_per_prompt: Optional[int] = 1,
|
796 |
+
eta: float = 0.0,
|
797 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
798 |
+
latents: Optional[torch.Tensor] = None,
|
799 |
+
prompt_embeds: Optional[torch.Tensor] = None,
|
800 |
+
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
801 |
+
ip_adapter_image: Optional[PipelineImageInput] = None,
|
802 |
+
ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
|
803 |
+
output_type: Optional[str] = "pil",
|
804 |
+
return_dict: bool = True,
|
805 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
806 |
+
guidance_rescale: float = 0.0,
|
807 |
+
clip_skip: Optional[int] = None,
|
808 |
+
callback_on_step_end: Optional[
|
809 |
+
Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
|
810 |
+
] = None,
|
811 |
+
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
812 |
+
**kwargs,
|
813 |
+
):
|
814 |
+
r"""
|
815 |
+
The call function to the pipeline for generation.
|
816 |
+
|
817 |
+
Args:
|
818 |
+
prompt (`str` or `List[str]`, *optional*):
|
819 |
+
The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
|
820 |
+
height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
|
821 |
+
The height in pixels of the generated image.
|
822 |
+
width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
|
823 |
+
The width in pixels of the generated image.
|
824 |
+
num_inference_steps (`int`, *optional*, defaults to 50):
|
825 |
+
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
826 |
+
expense of slower inference.
|
827 |
+
timesteps (`List[int]`, *optional*):
|
828 |
+
Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
|
829 |
+
in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
|
830 |
+
passed will be used. Must be in descending order.
|
831 |
+
sigmas (`List[float]`, *optional*):
|
832 |
+
Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
|
833 |
+
their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
|
834 |
+
will be used.
|
835 |
+
guidance_scale (`float`, *optional*, defaults to 7.5):
|
836 |
+
A higher guidance scale value encourages the model to generate images closely linked to the text
|
837 |
+
`prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
|
838 |
+
negative_prompt (`str` or `List[str]`, *optional*):
|
839 |
+
The prompt or prompts to guide what to not include in image generation. If not defined, you need to
|
840 |
+
pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
|
841 |
+
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
842 |
+
The number of images to generate per prompt.
|
843 |
+
eta (`float`, *optional*, defaults to 0.0):
|
844 |
+
Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
|
845 |
+
to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
|
846 |
+
generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
|
847 |
+
A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
|
848 |
+
generation deterministic.
|
849 |
+
latents (`torch.Tensor`, *optional*):
|
850 |
+
Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
|
851 |
+
generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
|
852 |
+
tensor is generated by sampling using the supplied random `generator`.
|
853 |
+
prompt_embeds (`torch.Tensor`, *optional*):
|
854 |
+
Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
|
855 |
+
provided, text embeddings are generated from the `prompt` input argument.
|
856 |
+
negative_prompt_embeds (`torch.Tensor`, *optional*):
|
857 |
+
Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
|
858 |
+
not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
|
859 |
+
ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
|
860 |
+
ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
|
861 |
+
Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
|
862 |
+
IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should
|
863 |
+
contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not
|
864 |
+
provided, embeddings are computed from the `ip_adapter_image` input argument.
|
865 |
+
output_type (`str`, *optional*, defaults to `"pil"`):
|
866 |
+
The output format of the generated image. Choose between `PIL.Image` or `np.array`.
|
867 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
868 |
+
Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
|
869 |
+
plain tuple.
|
870 |
+
cross_attention_kwargs (`dict`, *optional*):
|
871 |
+
A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
|
872 |
+
[`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
873 |
+
guidance_rescale (`float`, *optional*, defaults to 0.0):
|
874 |
+
Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
|
875 |
+
Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
|
876 |
+
using zero terminal SNR.
|
877 |
+
clip_skip (`int`, *optional*):
|
878 |
+
Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
|
879 |
+
the output of the pre-final layer will be used for computing the prompt embeddings.
|
880 |
+
callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
|
881 |
+
A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
|
882 |
+
each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
|
883 |
+
DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
|
884 |
+
list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
|
885 |
+
callback_on_step_end_tensor_inputs (`List`, *optional*):
|
886 |
+
The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
|
887 |
+
will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
|
888 |
+
`._callback_tensor_inputs` attribute of your pipeline class.
|
889 |
+
|
890 |
+
Examples:
|
891 |
+
|
892 |
+
Returns:
|
893 |
+
[`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
|
894 |
+
If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
|
895 |
+
otherwise a `tuple` is returned where the first element is a list with the generated images and the
|
896 |
+
second element is a list of `bool`s indicating whether the corresponding generated image contains
|
897 |
+
"not-safe-for-work" (nsfw) content.
|
898 |
+
"""
|
899 |
+
|
900 |
+
callback = kwargs.pop("callback", None)
|
901 |
+
callback_steps = kwargs.pop("callback_steps", None)
|
902 |
+
self.model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
|
903 |
+
self._optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
|
904 |
+
self._exclude_from_cpu_offload = ["safety_checker"]
|
905 |
+
self._callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
|
906 |
+
|
907 |
+
if callback is not None:
|
908 |
+
deprecate(
|
909 |
+
"callback",
|
910 |
+
"1.0.0",
|
911 |
+
"Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
|
912 |
+
)
|
913 |
+
if callback_steps is not None:
|
914 |
+
deprecate(
|
915 |
+
"callback_steps",
|
916 |
+
"1.0.0",
|
917 |
+
"Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
|
918 |
+
)
|
919 |
+
|
920 |
+
if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
|
921 |
+
callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
|
922 |
+
|
923 |
+
# 0. Default height and width to unet
|
924 |
+
if not height or not width:
|
925 |
+
height = (
|
926 |
+
self.unet.config.sample_size
|
927 |
+
if self._is_unet_config_sample_size_int
|
928 |
+
else self.unet.config.sample_size[0]
|
929 |
+
)
|
930 |
+
width = (
|
931 |
+
self.unet.config.sample_size
|
932 |
+
if self._is_unet_config_sample_size_int
|
933 |
+
else self.unet.config.sample_size[1]
|
934 |
+
)
|
935 |
+
height, width = height * self.vae_scale_factor, width * self.vae_scale_factor
|
936 |
+
# to deal with lora scaling and other possible forward hooks
|
937 |
+
|
938 |
+
# 1. Check inputs. Raise error if not correct
|
939 |
+
self.check_inputs(
|
940 |
+
prompt,
|
941 |
+
height,
|
942 |
+
width,
|
943 |
+
callback_steps,
|
944 |
+
negative_prompt,
|
945 |
+
prompt_embeds,
|
946 |
+
negative_prompt_embeds,
|
947 |
+
ip_adapter_image,
|
948 |
+
ip_adapter_image_embeds,
|
949 |
+
callback_on_step_end_tensor_inputs,
|
950 |
+
)
|
951 |
+
|
952 |
+
self._guidance_scale = guidance_scale
|
953 |
+
self._guidance_rescale = guidance_rescale
|
954 |
+
self._clip_skip = clip_skip
|
955 |
+
self._cross_attention_kwargs = cross_attention_kwargs
|
956 |
+
self._interrupt = False
|
957 |
+
|
958 |
+
# 2. Define call parameters
|
959 |
+
if prompt is not None and isinstance(prompt, str):
|
960 |
+
batch_size = 1
|
961 |
+
elif prompt is not None and isinstance(prompt, list):
|
962 |
+
batch_size = len(prompt)
|
963 |
+
else:
|
964 |
+
batch_size = prompt_embeds.shape[0]
|
965 |
+
|
966 |
+
device = self._execution_device
|
967 |
+
|
968 |
+
# 3. Encode input prompt
|
969 |
+
lora_scale = (
|
970 |
+
self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
|
971 |
+
)
|
972 |
+
|
973 |
+
prompt_embeds, negative_prompt_embeds = self.encode_prompt(
|
974 |
+
prompt,
|
975 |
+
device,
|
976 |
+
num_images_per_prompt,
|
977 |
+
self.do_classifier_free_guidance,
|
978 |
+
negative_prompt,
|
979 |
+
prompt_embeds=prompt_embeds,
|
980 |
+
negative_prompt_embeds=negative_prompt_embeds,
|
981 |
+
lora_scale=lora_scale,
|
982 |
+
clip_skip=self.clip_skip,
|
983 |
+
)
|
984 |
+
|
985 |
+
# For classifier free guidance, we need to do two forward passes.
|
986 |
+
# Here we concatenate the unconditional and text embeddings into a single batch
|
987 |
+
# to avoid doing two forward passes
|
988 |
+
if self.do_classifier_free_guidance:
|
989 |
+
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
|
990 |
+
|
991 |
+
if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
|
992 |
+
image_embeds = self.prepare_ip_adapter_image_embeds(
|
993 |
+
ip_adapter_image,
|
994 |
+
ip_adapter_image_embeds,
|
995 |
+
device,
|
996 |
+
batch_size * num_images_per_prompt,
|
997 |
+
self.do_classifier_free_guidance,
|
998 |
+
)
|
999 |
+
|
1000 |
+
# 4. Prepare timesteps
|
1001 |
+
timesteps, num_inference_steps = retrieve_timesteps(
|
1002 |
+
self.scheduler, num_inference_steps, device, timesteps, sigmas
|
1003 |
+
)
|
1004 |
+
|
1005 |
+
# 5. Prepare latent variables
|
1006 |
+
num_channels_latents = self.unet.config.in_channels
|
1007 |
+
latents = self.prepare_latents(
|
1008 |
+
batch_size * num_images_per_prompt,
|
1009 |
+
num_channels_latents,
|
1010 |
+
height,
|
1011 |
+
width,
|
1012 |
+
prompt_embeds.dtype,
|
1013 |
+
device,
|
1014 |
+
generator,
|
1015 |
+
latents,
|
1016 |
+
)
|
1017 |
+
|
1018 |
+
# 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
|
1019 |
+
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
1020 |
+
|
1021 |
+
# 6.1 Add image embeds for IP-Adapter
|
1022 |
+
added_cond_kwargs = (
|
1023 |
+
{"image_embeds": image_embeds}
|
1024 |
+
if (ip_adapter_image is not None or ip_adapter_image_embeds is not None)
|
1025 |
+
else None
|
1026 |
+
)
|
1027 |
+
|
1028 |
+
# 6.2 Optionally get Guidance Scale Embedding
|
1029 |
+
timestep_cond = None
|
1030 |
+
if self.unet.config.time_cond_proj_dim is not None:
|
1031 |
+
guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
|
1032 |
+
timestep_cond = self.get_guidance_scale_embedding(
|
1033 |
+
guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
|
1034 |
+
).to(device=device, dtype=latents.dtype)
|
1035 |
+
|
1036 |
+
# 7. Denoising loop
|
1037 |
+
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
|
1038 |
+
self._num_timesteps = len(timesteps)
|
1039 |
+
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
1040 |
+
for i, t in enumerate(timesteps):
|
1041 |
+
if self.interrupt:
|
1042 |
+
continue
|
1043 |
+
|
1044 |
+
# expand the latents if we are doing classifier free guidance
|
1045 |
+
latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
|
1046 |
+
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
|
1047 |
+
|
1048 |
+
# predict the noise residual
|
1049 |
+
noise_pred = self.unet(
|
1050 |
+
latent_model_input,
|
1051 |
+
t,
|
1052 |
+
encoder_hidden_states=prompt_embeds,
|
1053 |
+
timestep_cond=timestep_cond,
|
1054 |
+
cross_attention_kwargs=self.cross_attention_kwargs,
|
1055 |
+
added_cond_kwargs=added_cond_kwargs,
|
1056 |
+
return_dict=False,
|
1057 |
+
)[0]
|
1058 |
+
|
1059 |
+
# perform guidance
|
1060 |
+
if self.do_classifier_free_guidance:
|
1061 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
1062 |
+
noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
|
1063 |
+
|
1064 |
+
if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
|
1065 |
+
# Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
|
1066 |
+
noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
|
1067 |
+
|
1068 |
+
# compute the previous noisy sample x_t -> x_t-1
|
1069 |
+
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
|
1070 |
+
|
1071 |
+
if callback_on_step_end is not None:
|
1072 |
+
callback_kwargs = {}
|
1073 |
+
for k in callback_on_step_end_tensor_inputs:
|
1074 |
+
callback_kwargs[k] = locals()[k]
|
1075 |
+
callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
|
1076 |
+
|
1077 |
+
latents = callback_outputs.pop("latents", latents)
|
1078 |
+
prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
|
1079 |
+
negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
|
1080 |
+
|
1081 |
+
# call the callback, if provided
|
1082 |
+
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
1083 |
+
progress_bar.update()
|
1084 |
+
if callback is not None and i % callback_steps == 0:
|
1085 |
+
step_idx = i // getattr(self.scheduler, "order", 1)
|
1086 |
+
callback(step_idx, t, latents)
|
1087 |
+
|
1088 |
+
if XLA_AVAILABLE:
|
1089 |
+
xm.mark_step()
|
1090 |
+
|
1091 |
+
if not output_type == "latent":
|
1092 |
+
image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
|
1093 |
+
0
|
1094 |
+
]
|
1095 |
+
image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
|
1096 |
+
else:
|
1097 |
+
image = latents
|
1098 |
+
has_nsfw_concept = None
|
1099 |
+
|
1100 |
+
if has_nsfw_concept is None:
|
1101 |
+
do_denormalize = [True] * image.shape[0]
|
1102 |
+
else:
|
1103 |
+
do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
|
1104 |
+
image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
|
1105 |
+
|
1106 |
+
# Offload all models
|
1107 |
+
self.maybe_free_model_hooks()
|
1108 |
+
|
1109 |
+
if not return_dict:
|
1110 |
+
return (image, has_nsfw_concept)
|
1111 |
+
|
1112 |
+
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
|
1113 |
+
|
1114 |
+
# copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion
|
1115 |
+
def _encode_prompt(
|
1116 |
+
self,
|
1117 |
+
prompt,
|
1118 |
+
device,
|
1119 |
+
num_images_per_prompt,
|
1120 |
+
do_classifier_free_guidance,
|
1121 |
+
negative_prompt=None,
|
1122 |
+
prompt_embeds: Optional[torch.Tensor] = None,
|
1123 |
+
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
1124 |
+
lora_scale: Optional[float] = None,
|
1125 |
+
**kwargs,
|
1126 |
+
):
|
1127 |
+
r"""Encodes the prompt into text encoder hidden states."""
|
1128 |
+
batch_size = len(prompt) if isinstance(prompt, list) else 1
|
1129 |
+
|
1130 |
+
# get prompt text embeddings
|
1131 |
+
text_inputs = self.tokenizer(
|
1132 |
+
prompt,
|
1133 |
+
padding="max_length",
|
1134 |
+
max_length=self.tokenizer.model_max_length,
|
1135 |
+
truncation=True,
|
1136 |
+
return_tensors="pt",
|
1137 |
+
)
|
1138 |
+
text_input_ids = text_inputs.input_ids
|
1139 |
+
untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
|
1140 |
+
|
1141 |
+
if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
|
1142 |
+
removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
|
1143 |
+
logger.warning(
|
1144 |
+
"The following part of your input was truncated because CLIP can only handle sequences up to"
|
1145 |
+
f" {self.tokenizer.model_max_length} tokens: {removed_text}"
|
1146 |
+
)
|
1147 |
+
|
1148 |
+
if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
|
1149 |
+
attention_mask = text_inputs.attention_mask.to(device)
|
1150 |
+
else:
|
1151 |
+
attention_mask = None
|
1152 |
+
|
1153 |
+
if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND:
|
1154 |
+
# cast text_encoder.dtype to prevent overflow when using bf16
|
1155 |
+
text_input_ids = text_input_ids.to(device=device, dtype=self.text_encoder.dtype)
|
1156 |
+
prompt_embeds = self.text_encoder(
|
1157 |
+
text_input_ids,
|
1158 |
+
attention_mask=attention_mask,
|
1159 |
+
)
|
1160 |
+
prompt_embeds = prompt_embeds[0]
|
1161 |
+
else:
|
1162 |
+
text_encoder_lora_scale = None
|
1163 |
+
if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin):
|
1164 |
+
text_encoder_lora_scale = lora_scale
|
1165 |
+
if text_encoder_lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin):
|
1166 |
+
# dynamically adjust the LoRA scale
|
1167 |
+
adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
|
1168 |
+
|
1169 |
+
prompt_embeds = self.text_encoder(
|
1170 |
+
text_input_ids.to(device),
|
1171 |
+
attention_mask=attention_mask,
|
1172 |
+
)
|
1173 |
+
prompt_embeds = prompt_embeds[0]
|
1174 |
+
|
1175 |
+
# duplicate text embeddings for each generation per prompt
|
1176 |
+
bs_embed, seq_len, _ = prompt_embeds.shape
|
1177 |
+
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
1178 |
+
prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
|
1179 |
+
|
1180 |
+
# get unconditional embeddings for classifier free guidance
|
1181 |
+
if do_classifier_free_guidance and negative_prompt_embeds is None:
|
1182 |
+
uncond_tokens: List[str]
|
1183 |
+
if negative_prompt is None:
|
1184 |
+
uncond_tokens = [""]
|
1185 |
+
elif type(prompt) is not type(negative_prompt):
|
1186 |
+
raise TypeError(
|
1187 |
+
f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
|
1188 |
+
f" {type(prompt)}."
|
1189 |
+
)
|
1190 |
+
elif isinstance(negative_prompt, str):
|
1191 |
+
uncond_tokens = [negative_prompt]
|
1192 |
+
elif batch_size != len(negative_prompt):
|
1193 |
+
raise ValueError(
|
1194 |
+
f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
|
1195 |
+
f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
|
1196 |
+
" the batch size of `prompt`."
|
1197 |
+
)
|
1198 |
+
else:
|
1199 |
+
uncond_tokens = negative_prompt
|
1200 |
+
|
1201 |
+
# textual inversion: process multi-vector tokens if necessary
|
1202 |
+
if isinstance(self, TextualInversionLoaderMixin):
|
1203 |
+
uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
|
1204 |
+
|
1205 |
+
max_length = prompt_embeds.shape[1]
|
1206 |
+
uncond_input = self.tokenizer(
|
1207 |
+
uncond_tokens,
|
1208 |
+
padding="max_length",
|
1209 |
+
max_length=max_length,
|
1210 |
+
truncation=True,
|
1211 |
+
return_tensors="pt",
|
1212 |
+
)
|
1213 |
+
|
1214 |
+
if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
|
1215 |
+
attention_mask = uncond_input.attention_mask.to(device)
|
1216 |
+
else:
|
1217 |
+
attention_mask = None
|
1218 |
+
|
1219 |
+
negative_prompt_embeds = self.text_encoder(
|
1220 |
+
uncond_input.input_ids.to(device),
|
1221 |
+
attention_mask=attention_mask,
|
1222 |
+
)
|
1223 |
+
negative_prompt_embeds = negative_prompt_embeds[0]
|
1224 |
+
|
1225 |
+
if do_classifier_free_guidance:
|
1226 |
+
# duplicate unconditional embeddings for each generation per prompt
|
1227 |
+
seq_len = negative_prompt_embeds.shape[1]
|
1228 |
+
|
1229 |
+
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
1230 |
+
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
|
1231 |
+
|
1232 |
+
if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND:
|
1233 |
+
# Unscale LoRA weights to avoid overfitting. This is a hack
|
1234 |
+
unscale_lora_layers(self.text_encoder, lora_scale)
|
1235 |
+
|
1236 |
+
return prompt_embeds, negative_prompt_embeds
|
1237 |
+
|
1238 |
+
def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
|
1239 |
+
"""Encodes the image into image encoder hidden states."""
|
1240 |
+
dtype = next(self.image_encoder.parameters()).dtype
|
1241 |
+
|
1242 |
+
if not isinstance(image, torch.Tensor):
|
1243 |
+
image = self.feature_extractor(image, return_tensors="pt").pixel_values
|
1244 |
+
|
1245 |
+
image = image.to(device=device, dtype=dtype)
|
1246 |
+
if output_hidden_states:
|
1247 |
+
image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
|
1248 |
+
image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
|
1249 |
+
uncond_image_enc_hidden_states = self.image_encoder(
|
1250 |
+
torch.zeros_like(image), output_hidden_states=True
|
1251 |
+
).hidden_states[-2]
|
1252 |
+
uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
|
1253 |
+
num_images_per_prompt, dim=0
|
1254 |
+
)
|
1255 |
+
return image_enc_hidden_states, uncond_image_enc_hidden_states
|
1256 |
+
else:
|
1257 |
+
image_embeds = self.image_encoder(image).image_embeds
|
1258 |
+
image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
|
1259 |
+
uncond_image_embeds = torch.zeros_like(image_embeds)
|
1260 |
+
|
1261 |
+
return image_embeds, uncond_image_embeds
|
1262 |
+
|
1263 |
+
def prepare_ip_adapter_image_embeds(
|
1264 |
+
self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
|
1265 |
+
):
|
1266 |
+
"""Prepares and processes IP-Adapter image embeddings."""
|
1267 |
+
image_embeds = []
|
1268 |
+
if do_classifier_free_guidance:
|
1269 |
+
negative_image_embeds = []
|
1270 |
+
if ip_adapter_image_embeds is None:
|
1271 |
+
for image in ip_adapter_image:
|
1272 |
+
if not isinstance(image, torch.Tensor):
|
1273 |
+
image = self.image_processor.preprocess(image)
|
1274 |
+
image = image.to(device=device)
|
1275 |
+
if len(image.shape) == 3:
|
1276 |
+
image = image.unsqueeze(0)
|
1277 |
+
image_emb, neg_image_emb = self.encode_image(image, device, num_images_per_prompt, True)
|
1278 |
+
image_embeds.append(image_emb)
|
1279 |
+
if do_classifier_free_guidance:
|
1280 |
+
negative_image_embeds.append(neg_image_emb)
|
1281 |
+
|
1282 |
+
if len(image_embeds) == 1:
|
1283 |
+
image_embeds = image_embeds[0]
|
1284 |
+
if do_classifier_free_guidance:
|
1285 |
+
negative_image_embeds = negative_image_embeds[0]
|
1286 |
+
else:
|
1287 |
+
image_embeds = torch.cat(image_embeds, dim=0)
|
1288 |
+
if do_classifier_free_guidance:
|
1289 |
+
negative_image_embeds = torch.cat(negative_image_embeds, dim=0)
|
1290 |
+
else:
|
1291 |
+
repeat_dim = 2 if do_classifier_free_guidance else 1
|
1292 |
+
image_embeds = ip_adapter_image_embeds.repeat_interleave(repeat_dim, dim=0)
|
1293 |
+
if do_classifier_free_guidance:
|
1294 |
+
negative_image_embeds = torch.zeros_like(image_embeds)
|
1295 |
+
|
1296 |
+
if do_classifier_free_guidance:
|
1297 |
+
image_embeds = torch.cat([negative_image_embeds, image_embeds])
|
1298 |
+
|
1299 |
+
return image_embeds
|
1300 |
+
|
1301 |
+
def run_safety_checker(self, image, device, dtype):
|
1302 |
+
"""Runs the safety checker on the generated image."""
|
1303 |
+
if self.safety_checker is None:
|
1304 |
+
has_nsfw_concept = None
|
1305 |
+
return image, has_nsfw_concept
|
1306 |
+
|
1307 |
+
if isinstance(self.safety_checker, StableDiffusionSafetyChecker):
|
1308 |
+
safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
|
1309 |
+
image, has_nsfw_concept = self.safety_checker(
|
1310 |
+
images=image,
|
1311 |
+
clip_input=safety_checker_input.pixel_values.to(dtype),
|
1312 |
+
)
|
1313 |
+
else:
|
1314 |
+
images_np = self.numpy_to_pil(image)
|
1315 |
+
safety_checker_input = self.safety_checker.feature_extractor(images_np, return_tensors="pt").to(device)
|
1316 |
+
has_nsfw_concept = self.safety_checker(
|
1317 |
+
images=image,
|
1318 |
+
clip_input=safety_checker_input.pixel_values.to(dtype),
|
1319 |
+
)[1]
|
1320 |
+
|
1321 |
+
return image, has_nsfw_concept
|
1322 |
+
|
1323 |
+
# copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion
|
1324 |
+
def decode_latents(self, latents):
|
1325 |
+
"""Decodes the latents to images."""
|
1326 |
+
latents = 1 / self.vae.config.scaling_factor * latents
|
1327 |
+
image = self.vae.decode(latents, return_dict=False)[0]
|
1328 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
1329 |
+
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
|
1330 |
+
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
1331 |
+
return image
|
1332 |
+
|
1333 |
+
@property
|
1334 |
+
def guidance_scale(self):
|
1335 |
+
return self._guidance_scale
|
1336 |
+
|
1337 |
+
@property
|
1338 |
+
def guidance_rescale(self):
|
1339 |
+
return self._guidance_rescale
|
1340 |
+
|
1341 |
+
# copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion
|
1342 |
+
def get_guidance_scale_embedding(
|
1343 |
+
self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32
|
1344 |
+
):
|
1345 |
+
"""Gets the guidance scale embedding for classifier free guidance conditioning.
|
1346 |
+
See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
|
1347 |
+
|
1348 |
+
Args:
|
1349 |
+
w (`torch.Tensor`):
|
1350 |
+
The guidance scale tensor used for classifier free guidance conditioning.
|
1351 |
+
embedding_dim (`int`, defaults to 512):
|
1352 |
+
The dimensionality of the guidance scale embedding.
|
1353 |
+
dtype (`torch.dtype`, defaults to torch.float32):
|
1354 |
+
The dtype of the embedding.
|
1355 |
+
|
1356 |
+
Returns:
|
1357 |
+
`torch.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`.
|
1358 |
+
"""
|
1359 |
+
assert len(w.shape) == 1
|
1360 |
+
w = w * 1000.0
|
1361 |
+
|
1362 |
+
half_dim = embedding_dim // 2
|
1363 |
+
emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
|
1364 |
+
emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
|
1365 |
+
emb = w.to(dtype)[:, None] * emb[None, :]
|
1366 |
+
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
|
1367 |
+
if embedding_dim % 2 == 1: # zero pad
|
1368 |
+
emb = torch.nn.functional.pad(emb, (0, 1))
|
1369 |
+
assert emb.shape == (w.shape[0], embedding_dim)
|
1370 |
+
return emb
|
1371 |
+
|
1372 |
+
@property
|
1373 |
+
def clip_skip(self):
|
1374 |
+
return self._clip_skip
|
1375 |
+
|
1376 |
+
@property
|
1377 |
+
def num_timesteps(self):
|
1378 |
+
return self._num_timesteps
|
1379 |
+
|
1380 |
+
@property
|
1381 |
+
def interrupt(self):
|
1382 |
+
return self._interrupt
|
1383 |
+
|
1384 |
+
@property
|
1385 |
+
def cross_attention_kwargs(self):
|
1386 |
+
return self._cross_attention_kwargs
|
1387 |
+
|
1388 |
+
@property
|
1389 |
+
def do_classifier_free_guidance(self):
|
1390 |
+
return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
|
1391 |
+
|
1392 |
|
1393 |
### Make prompt list for each regions
|
1394 |
def promptsmaker(prompts, batch):
|
|
|
1603 |
get_attn_maps(self, attn_weight)
|
1604 |
attn_weight = torch.dropout(attn_weight, dropout_p, train=True)
|
1605 |
return attn_weight @ value
|
1606 |
+
|
1607 |
+
|
1608 |
+
def retrieve_timesteps(
|
1609 |
+
scheduler,
|
1610 |
+
num_inference_steps: Optional[int] = None,
|
1611 |
+
device: Optional[Union[str, torch.device]] = None,
|
1612 |
+
timesteps: Optional[List[int]] = None,
|
1613 |
+
sigmas: Optional[List[float]] = None,
|
1614 |
+
**kwargs,
|
1615 |
+
):
|
1616 |
+
r"""
|
1617 |
+
Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
|
1618 |
+
custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
|
1619 |
+
|
1620 |
+
Args:
|
1621 |
+
scheduler (`SchedulerMixin`):
|
1622 |
+
The scheduler to get timesteps from.
|
1623 |
+
num_inference_steps (`int`):
|
1624 |
+
The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
|
1625 |
+
must be `None`.
|
1626 |
+
device (`str` or `torch.device`, *optional*):
|
1627 |
+
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
|
1628 |
+
timesteps (`List[int]`, *optional*):
|
1629 |
+
Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
|
1630 |
+
`num_inference_steps` and `sigmas` must be `None`.
|
1631 |
+
sigmas (`List[float]`, *optional*):
|
1632 |
+
Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
|
1633 |
+
`num_inference_steps` and `timesteps` must be `None`.
|
1634 |
+
|
1635 |
+
Returns:
|
1636 |
+
`Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
|
1637 |
+
second element is the number of inference steps.
|
1638 |
+
"""
|
1639 |
+
if timesteps is not None and sigmas is not None:
|
1640 |
+
raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
|
1641 |
+
if timesteps is not None:
|
1642 |
+
accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
|
1643 |
+
if not accepts_timesteps:
|
1644 |
+
raise ValueError(
|
1645 |
+
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
|
1646 |
+
f" timestep schedules. Please check whether you are using the correct scheduler."
|
1647 |
+
)
|
1648 |
+
scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
|
1649 |
+
timesteps = scheduler.timesteps
|
1650 |
+
num_inference_steps = len(timesteps)
|
1651 |
+
elif sigmas is not None:
|
1652 |
+
accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
|
1653 |
+
if not accept_sigmas:
|
1654 |
+
raise ValueError(
|
1655 |
+
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
|
1656 |
+
f" sigmas schedules. Please check whether you are using the correct scheduler."
|
1657 |
+
)
|
1658 |
+
scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
|
1659 |
+
timesteps = scheduler.timesteps
|
1660 |
+
num_inference_steps = len(timesteps)
|
1661 |
+
else:
|
1662 |
+
scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
|
1663 |
+
timesteps = scheduler.timesteps
|
1664 |
+
return timesteps, num_inference_steps
|
1665 |
+
|
1666 |
+
|
1667 |
+
def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
|
1668 |
+
r"""
|
1669 |
+
Rescales `noise_cfg` tensor based on `guidance_rescale` to improve image quality and fix overexposure. Based on
|
1670 |
+
Section 3.4 from [Common Diffusion Noise Schedules and Sample Steps are
|
1671 |
+
Flawed](https://arxiv.org/pdf/2305.08891.pdf).
|
1672 |
+
|
1673 |
+
Args:
|
1674 |
+
noise_cfg (`torch.Tensor`):
|
1675 |
+
The predicted noise tensor for the guided diffusion process.
|
1676 |
+
noise_pred_text (`torch.Tensor`):
|
1677 |
+
The predicted noise tensor for the text-guided diffusion process.
|
1678 |
+
guidance_rescale (`float`, *optional*, defaults to 0.0):
|
1679 |
+
A rescale factor applied to the noise predictions.
|
1680 |
+
|
1681 |
+
Returns:
|
1682 |
+
noise_cfg (`torch.Tensor`): The rescaled noise prediction tensor.
|
1683 |
+
"""
|
1684 |
+
std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
|
1685 |
+
std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
|
1686 |
+
# rescale the results from guidance (fixes overexposure)
|
1687 |
+
noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
|
1688 |
+
# mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
|
1689 |
+
noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
|
1690 |
+
return noise_cfg
|